Cherry Blossom Peak Bloom Prediction 2026

5 Competition sites

3 Auxiliary sources: Japan regional (6,573), MeteoSwiss (6,642), South Korea (994)

USA-NPN enrichment — 5 extra NYC records from Washington Square Park citizen-science observations (species 228, phenophase 501)

Total: 14,598 bloom records

Enhanced features (from master dataset + NOAA CDO)

Model A — Local Trend (per site)

• Recency-weighted quadratic: bloom_doy ~ year + year²
• Exponential decay weights: \(w_i = e^{(i-n)/6}\), half-life ≈ 6 yr
• Captures site-specific momentum from long local histories

Model B — Pooled GAM (R pipeline)

\[\text{DOY} \sim s(\text{year}) + s(\text{lat}, \text{long}) + s(\text{alt}) + s(\text{obs}) + \text{climate} + \text{source}\]

• REML on 14,598 records
• NOAA CDO covariates: winter/spring temp, GDD proxy

Model B′ — Gradient Boosted Trees (Python pipeline)

• GBR with Huber loss, 800 trees, lr = 0.015
• Enhanced features: GDD, chill hours, ONI/NAO/PDO, Hopkins index
• Stochastic boosting (subsample=0.8, max_features=√p)

Ensemble blending

• Rolling-origin backtest (1900–2025, ~126 windows)
• Site-specific optimal weights via grid search (w ∈ [0, 1] by 0.02)
• Sites with deep history → heavier Model A; sparse sites → Model B

Prediction intervals

• Split-conformal: 90th percentile of backtest |residuals| per site
• ≥90% empirical coverage, tight enough for SSW tiebreaker

Cross-language guard

• R (GAM) and Python (GBR) run independently
• If avg gap ≤ 4 days → blend both; else default to R
• Current gap: 2.8 days → blended

Backtest MAE comparison

Key improvements:

• GDD + chill hours capture phenological drivers directly
• Climate indices (ONI, NAO, PDO) explain year-to-year variability
• Stochastic boosting + deeper trees capture complex interactions
• Site-specific weights outperform global inverse-MAE

Global-scale drivers

1. Rising temperatures — 1.1 °C warming since pre-industrial. Warmer winters reduce chill-hour needs; warmer springs accelerate GDD → bloom advances everywhere.

2. ENSO — La Niña cools Pacific NW (delays Vancouver), milder Eastern US winters (advances DC/NYC). 2025-26 neutral transition → near-normal timing.

3. Latitude gradient — ~1 DOY delay per degree N above 35°N. GAM’s lat-long smooth captures this.

4. Altitude penalty — ~2 days later per 100 m elevation. Liestal (350 m) systematically later than sea-level sites.

Site-specific drivers

• Kyoto: Urban heat island + 0.2 d/century advancement. 1200-yr record stabilises local model.

• Washington DC: Tidal Basin thermal buffer → earliest bloomer. 2.2 d/decade recent acceleration.

• Liestal: Alpine amplified warming (0.3 °C/decade). Foehn winds cause stochastic early bloom → widest variability.

• Vancouver: PDO phase modulates decadal variability. Maritime buffering keeps volatility moderate.

• NYC: Shortest record (12 yrs) → highest uncertainty. Continental cold snaps reset GDD accumulation. NPN data fusion reduces LOO MAE by ~1.5 days.

What the enhanced model adds:

• GDD + chill hours → direct phenological drivers
• ONI / NAO / PDO → macro-climate variability
• Hopkins index + photoperiod → bioclimatic baselines
• Deeper trees (depth 4) capture feature interactions
• Stochastic boosting reduces overfitting

Why enhanced predictions are later:

• Better calibration of spring warming accumulation
• Chill-hour accounting prevents premature bloom calls
• Climate indices correctly dampen warm-bias years

Tighter intervals (12.2 d avg vs 15.4 d baseline):

• More features → lower residual variance
• Better site-specific interval calibration
• Exception: Vancouver (25 d) — only 4 years of data